Optical neural network convolution computing chip device

Abstract

The application relates to a chip device for optical neural network convolution calculation, comprising a chip substrate, two or more light sources integrated on the substrate, a first wavelength division multiplexer, two or more first micro-ring cavity groups, a star coupler, a second wavelength division multiplexer and two or more detectors; the light output ends of the two or more light sources are connected with the input ends of the first wavelength division multiplexer respectively, the output end of the first wavelength division multiplexer is connected with the two or more first micro-ring cavity groups; the output ends of the two or more first micro-ring cavity groups are connected with the input ends of the star coupler respectively; the output ends of the star coupler are connected with the input end of the second wavelength division multiplexer through a second micro-ring cavity group respectively, and the output ends of the second wavelength division multiplexer are connected with the input ends of the two or more detectors, which are suitable for rough demultiplexing of optical signals and power calculation of the detectors.

Description

Technical Field

[0001] This application relates to the field of optical technology, and in particular to a chip device for optical neural network convolution calculation. Background Technology

[0002] In recent years, with the rise and development of optical neural networks and optical computing, optical chips have the characteristics of being able to directly complete information processing and convolution calculations on-chip, and have ultra-low power consumption, which provides a new approach to breaking through the bottlenecks of traditional computing.

[0003] Currently, the increase in computational complexity of optical neural network architectures mainly relies on the stacking of individual optical devices. As the number of optical devices and photonic neurons in the architecture increases, the achievable computational complexity also increases. However, the more optical devices are stacked, the more difficult it becomes to efficiently utilize their resources. During task processing, some optical devices may remain inactive, leading to significant resource waste. For example, micro-ring resonator arrays on a chip are an important component of photonic convolutional neural networks and a highly advantageous and widely studied physical structure for optical convolution computation. However, since each micro-ring resonator array corresponds to only one type of modulation information, when the optical signal to be convolved contains multiple modulation information, multiple micro-ring resonator arrays are needed to perform convolution processing on it separately. But when the optical signal input to the micro-ring resonator array contains only one type of modulation information, only one set of micro-ring resonator arrays processes it, leaving the remaining micro-ring resonator arrays on the chip idle, and their wavelength channels cannot be effectively utilized. This uneven resource allocation causes idle loss in photonic computing units.

[0004] Therefore, there is an urgent need for an optical neural network architecture that can reduce resource consumption. Summary of the Invention

[0005] In view of this, this application proposes a chip device for optical neural network convolution calculation.

[0006] According to one aspect of this application, a chip device for optical neural network convolution calculation is provided, comprising: a chip substrate, two or more light sources, a first wavelength division multiplexer, two or more first microring resonant cavity groups, a star coupler, two or more second microring resonant cavity groups, a second wavelength division multiplexer, and two or more detectors. All light sources, the first wavelength division multiplexer, all first microring resonator groups, star couplers, all second microring resonator groups, the second wavelength division multiplexer, and all detectors are integrated on the chip substrate. The optical output terminals of two or more light sources are respectively connected to the input terminals of the first wavelength division multiplexer to transmit multiple optical signals of different wavelengths to the first wavelength division multiplexer. The output terminal of the first wavelength division multiplexer is connected to two or more first micro-ring resonator groups to combine multiple optical signals and transmit them to each first micro-ring resonator group for modulation processing to obtain modulated optical signals. The outputs of two or more first micro-ring resonant cavity groups are respectively connected to the inputs of the star coupler; each output of the star coupler is connected to the input of the second wavelength division multiplexer through a second micro-ring resonant cavity group, so that when the modulated optical signal carries only one type of modulation information, the star coupler will divide the modulated optical signal into multiple paths and send them to each of the second micro-ring resonant cavity groups for simultaneous convolution processing. The outputs of the second wavelength division multiplexer are connected to the inputs of two or more detectors, which is suitable for coarsely demultiplexing optical signals and sending them to the detectors for power calculation. The number of microrings in each first microring resonator group and the number of microrings in each second microring resonator group are the same as the number of light sources.

[0007] In one possible implementation, the output of one of the first microring resonator groups is connected to one input of a star coupler, and the outputs of the remaining first microring resonator groups are connected to the remaining inputs of the star coupler via optical extension lines; and the lengths of the optical extension lines connected to each first microring resonator group increase or decrease sequentially.

[0008] In one possible implementation, the length difference of the optical extension line connecting any two adjacent first microring resonator groups is equal.

[0009] In one possible implementation, the light source is a laser.

[0010] In one possible implementation, the number of light sources is eight.

[0011] In one possible implementation, there are four first microring resonator groups, and each first microring resonator group contains eight microrings.

[0012] In one possible implementation, the number of second microring resonator groups is four.

[0013] In one possible implementation, the star coupler is a 4×4 coupler.

[0014] Beneficial effects: Two or more light sources are suitable for emitting optical signals of different wavelengths. The first wavelength division multiplexer is suitable for multiplexing and combining multiple optical signals of different wavelengths into a composite optical signal. Then, the composite optical signal is divided into multiple beams and transmitted to each of the first micro-ring resonator groups. All the first micro-ring resonator groups are suitable for modulating the received composite optical signal to obtain the modulated optical signal. The modulated optical signal is sent to each input terminal of the star coupler. The star coupler is suitable for channel switching and can distribute the modulated optical signal to each of the second micro-ring resonator groups, so that multiple second micro-ring resonator groups after the star coupler can receive the modulated optical signal. Even if the modulated optical signal carries only one modulation information, multiple second micro-ring resonator groups can simultaneously perform convolution processing on the modulated optical signal, which greatly improves resource utilization. The second wavelength division multiplexer is suitable for combining the power of adjacent different wavelengths and sending the combined optical signal to each detector. The detector detects the sum of the optical power of two adjacent wavelengths. This application integrates the three functions of transmission, processing, and detection onto a single chip substrate, which greatly improves the chip's integration level and avoids the waste of computing resources.

[0015] Other features and aspects of this application will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0016] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this application together with the specification and serve to explain the principles of this application.

[0017] Figure 1 This diagram shows the main structure of a chip device for optical neural network convolution calculation according to an embodiment of this application.

[0018] Chip substrate 100, light source 200, first wavelength division multiplexer 300, first micro-ring resonator group 400, star coupler 600, second micro-ring resonator group 700, second wavelength division multiplexer 800, detector 900. Detailed Implementation

[0019] Various exemplary embodiments, features, and aspects of the present invention will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0020] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this utility model or simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this utility model.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this utility model, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0023] Furthermore, to better illustrate this utility model, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that this utility model can be implemented even without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail, in order to highlight the main points of this utility model.

[0024] Figure 1 This diagram illustrates the main structure of a chip device for optical neural network convolution calculation according to an embodiment of this application. Figure 1As shown, this optical neural network convolution calculation chip device includes: a chip substrate 100, two or more light sources 200, a first wavelength division multiplexer 300, two or more first micro-ring resonator groups 400, a star coupler 600, two or more second micro-ring resonator groups 700, a second wavelength division multiplexer 800, and a detector 900; the two or more light sources 200, the first wavelength division multiplexer 300, the first micro-ring resonator group 400, the star coupler 600, the second micro-ring resonator group 700, the second wavelength division multiplexer 800, and the detector 900 are all integrated on the chip substrate 100; the optical output terminals of the two or more light sources 200 are respectively connected to the input terminals of the first wavelength division multiplexer 300 to transmit multiple optical signals of different wavelengths to the first wavelength division multiplexer 300; the output terminal of the first wavelength division multiplexer 300 is connected to the two or more first micro-ring resonator groups 400 to convert multiple optical signals... After beam combining, the signal is transmitted to each of the first microring resonator groups 400 for modulation processing to obtain the modulated optical signal. The outputs of two or more first microring resonator groups 400 are respectively connected to the inputs of star couplers 600. Each output of star couplers 600 is connected to the input of a second wavelength division multiplexer 800 through a second microring resonator group 700. When the modulated optical signal carries only one modulation information, the star coupler 600 distributes the modulated optical signal into multiple paths and sends them to each of the second microring resonator groups 400 for simultaneous convolution processing. Each output of the second wavelength division multiplexer 800 is connected to the input of two or more detectors 900 for coarse demultiplexing of the optical signal and sending it to the detectors 900 for power calculation. The number of microrings in each first microring resonator group 400 and the number of microrings in each second microring resonator group 700 are the same as the number of light sources 200.

[0025] It should be noted here that two or more light sources 200 are suitable for emitting optical signals of different wavelengths respectively, and the first wavelength division multiplexer 300 is suitable for multiplexing multiple optical signals of different wavelengths into a composite optical signal. Then the composite optical signal is divided into multiple beams and transmitted to each of the first micro-ring resonant cavity groups 400. These multiple composite optical signals contain signals of various different wavelengths. All first microring resonator groups 400 are suitable for modulating the received composite optical signal to obtain a modulated optical signal (a beam of light containing modulation information), and the modulated optical signal contains light of different wavelengths with different optical powers. The modulated optical signal is sent to each input terminal of the star coupler 600, which is suitable for channel switching. By reasonably distributing the thermally tuned electrodes on the star coupler 600, the optical path can be redistributed. The modulated optical signal can be distributed to each second microring resonator group 700, so that multiple second microring resonator groups 700 after the star coupler 600 can receive the modulated optical signal. Even if the modulated optical signal carries only one modulation information, multiple second microring resonator groups 700 can simultaneously perform convolution processing on the modulated optical signal, which greatly improves resource utilization. The second wavelength division multiplexer 300 is suitable for combining the power of two adjacent beams of different wavelength optical signals and sending the combined optical signal to each detector 900. The detector 900 detects the sum of the optical power of two adjacent beams of different wavelength optical signals. Since the lengths of the optical extension lines 500 connected to each of the first micro-ring resonator groups 400 are different and tend to increase or decrease sequentially, time-division multiplexing can be performed through optical delay lines to avoid the overlap of optical power of each optical signal. Compared with other optical neural network architectures, this application integrates the three functions of emission, processing and detection on a single chip substrate 100, which greatly improves the chip integration. At the same time, in terms of architecture design, time-division multiplexing is performed through optical delay lines and channel switching is performed through star couplers 600, which avoids the waste of computing resources. Finally, the second wavelength division multiplexer 800 combines two adjacent beams of light of a specific wavelength, which greatly improves the architecture's computing efficiency and integration.

[0026] Specific explanation: Currently, each microring resonator group corresponds to one type of modulation information. If only one type of modulation information needs to be input into the optical signal, only one microring resonator group needs to be modulated; the remaining microring resonator groups do not carry modulation information. If multiple modulation information needs to be input into the optical signal, multiple microring resonator groups need to be modulated separately. For example: Figure 1As shown, in front of the star coupler 600, there are four first microring resonator groups 400, each corresponding to a set of modulation information. When the four first microring resonator groups 400 modulate the composite optical signal differently, the resulting modulated optical signal carries four types of modulation information. If only one first microring resonator group 400 modulates the composite optical signal, the resulting modulated optical signal carries only one type of modulation information. Behind the star coupler 600, there are two second microring resonator groups 700, each corresponding to a set of modulation information. When the modulated optical signal carries only one type of modulation information, this signal will only be input to one second microring resonator group 700 for convolution processing. The remaining three second microring resonator groups 700 will be inactive, resulting in wasted resources. However, this application achieves optical path redistribution by adding a star coupler 600. Even if the light at the input end of the star coupler 600 carries only one type of modulation information, it can be distributed to the four-way second micro-ring resonator group 700 for convolution processing, making full use of the on-chip computing resources.

[0027] In one possible implementation, the output of one of the first microring resonator groups 400 is connected to one input of the star coupler 600, and the outputs of the other first microring resonator groups 400 are respectively connected to the other inputs of the star coupler 600 through optical extension lines 500; and the lengths of the optical extension lines 500 connected to each first microring resonator group 400 increase or decrease sequentially.

[0028] In one possible implementation, the length difference of the optical extension lines 500 connecting any two adjacent first microring resonator groups 400 is the same. It should be noted that the length difference of the optical extension lines 500 mainly depends on the detector bandwidth of the entire architecture. For example, if the bandwidth is 40 GHz, a delay of 25 ps is required, resulting in a length difference of 1875 μm.

[0029] In one possible implementation, the first microring resonator group 400 contains a total of eight microrings. Due to the resonant characteristics of the microrings, different microrings will only adjust the power of light of a specific wavelength. The eight microrings modulate eight different wavelengths of light respectively. By adjusting the thermotunable electrodes located in the microring resonator, the optical power can be controlled from 1 to 0. After passing through a set of microring resonators, a beam of light containing modulation information is obtained.

[0030] In one possible implementation, the light source 200 is a laser, and each laser outputs light of only one wavelength. The output of each light source 200 is connected to the first wavelength division multiplexer 300 via optical fiber.

[0031] Furthermore, there are eight light sources 200, meaning that the eight lasers can output light signals of eight different wavelengths.

[0032] In one possible implementation, there are four first microring resonator groups 400. It should be noted that the first microring resonator group 400 represents input information, such as a matrix. By adjusting the offset of the notch center of the microring, the optical power is tuned, thus imbuing the optical power with the matrix information.

[0033] In one possible implementation, the number of second microring resonator groups 700 is four. The second microring resonator group 700 primarily functions as a convolution kernel, a common algorithm in neural networks.

[0034] The output of the first wavelength division multiplexer 300 is directly connected to each of the first micro-ring resonator groups 400, each of the first micro-ring resonator groups 400 is connected to the star coupler 600, the star coupler 600 is connected to each of the second micro-ring resonator groups 700, each of the second micro-ring resonator groups 700 is connected to the second wavelength division multiplexer 800, and the second wavelength division multiplexer 800 is connected to each of the detectors 900 via optical waveguides on the chip substrate 100.

[0035] In one possible implementation, there are four detectors 900. It should also be noted that the second wavelength division multiplexer 300 is a coarse demultiplexer that combines two adjacent wavelengths of light into one beam. Since there are eight light sources 200, eight different wavelengths of light are output. After the second wavelength division multiplexer 300 combines two adjacent wavelengths of light into one beam, it finally outputs four beams, which are then sent to the four detectors 900 to detect the sum of the optical power of the two adjacent wavelengths of light.

[0036] In one possible implementation, the star coupler 600 is a 4×4 coupler. The star coupler 600 employs thermal tuning control. If the light at the input of the star coupler 600 carries only one type of modulation information, it is distributed into four paths and sent to four second micro-ring resonator groups 700 for simultaneous convolution processing. If the light at the input carries four types of modulation information, it is distributed to four second micro-ring resonator groups 700 for individual convolution processing. The star coupler 600 primarily functions as a channel redistribution mechanism, making rational use of subsequent device resources and avoiding waste of on-chip computing units.

[0037] In one possible implementation, the chip substrate 100 is made of silicon-on-insulator; the first microring resonator group 400 and the second microring resonator group 700 are made of silicon-on-insulator or silicon-based silicon dioxide.

[0038] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A chip device for optical neural network convolution computation, characterized by, include: A chip substrate, two or more light sources, a first wavelength division multiplexer, two or more first microring resonant cavity groups, a star coupler, two or more second microring resonant cavity groups, a second wavelength division multiplexer, and two or more detectors; All of the light sources, the first wavelength division multiplexer, all of the first micro-ring resonator groups, the star coupler, all of the second micro-ring resonator groups, the second wavelength division multiplexer, and all of the detectors are integrated on the chip substrate; The light output terminals of two or more light sources are respectively connected to the input terminals of the first wavelength division multiplexer to transmit multiple optical signals of different wavelengths to the first wavelength division multiplexer. The output terminal of the first wavelength division multiplexer is connected to two or more first micro-ring resonator groups to combine the multiple optical signals and transmit them to each of the first micro-ring resonator groups for modulation processing to obtain modulated optical signals. The outputs of two or more first microring resonator groups are respectively connected to each input of the star coupler; each output of the star coupler is respectively connected to the input of the second wavelength division multiplexer through a second microring resonator group, so that when the modulated optical signal carries only one modulation information, the star coupler divides the modulated optical signal into multiple paths and sends them to each of the second microring resonator groups for simultaneous convolution processing; Each output terminal of the second wavelength division multiplexer is connected to the input terminals of two or more of the detectors to coarsely demultiplex the optical signal and send it to the detector for power calculation; The number of microrings in each of the first microring resonant cavity groups and the number of microrings in each of the second microring resonant cavity groups are the same as the number of light sources.

2. The optical neural network convolution computing chip apparatus according to claim 1, wherein, One of the first microring resonator groups has its output terminal connected to one of the input terminals of the star coupler, and the output terminals of the remaining first microring resonator groups are connected to the remaining input terminals of the star coupler through optical extension lines; and the lengths of the optical extension lines connected to each of the first microring resonator groups increase or decrease sequentially.

3. The chip device for optical neural network convolution calculation according to claim 2, characterized in that, The length difference of the optical extension line connecting any two adjacent first microring resonant cavity groups is equal.

4. The chip device for optical neural network convolution calculation according to claim 1, characterized in that, The light source is a laser.

5. The chip device for optical neural network convolution calculation according to claim 4, characterized in that, The number of light sources is eight.

6. The chip device for optical neural network convolution calculation according to claim 5, characterized in that, There are four first microring resonator groups, and each first microring resonator group contains eight microrings.

7. The chip device for optical neural network convolution calculation according to claim 6, characterized in that, The second micro-ring resonator group has four components.

8. The chip device for optical neural network convolution calculation according to claim 7, characterized in that, The star coupler is a 4×4 coupler.

Images